![]()
Part I
Multimedia Mining and Classification
![]()
1 | Multimedia Duplicate Mining toward Knowledge Discovery |
| Xiaomeng Wu, Sebastien Poullot and Shin’ichi Satoh |
CONTENTS
1.1 Introduction
1.2 Selection Criterion of Duplicate Mining Methods
1.2.1 Exact Duplicate Mining
1.2.2 Near-Duplicate Mining
1.3 TV Commercial Mining for Sociological Analysis
1.3.1 Background
1.3.2 Temporal Recurrence Hashing Algorithm
1.3.3 Knowledge Discovery Based on CF Mining
1.4 News Story Retrieval and Threading
1.4.1 Background
1.4.2 StoryRank
1.4.3 Experimentation Evaluation
1.5 Indexing for Scaling up Video Miningi
1.5.1 Background
1.5.2 Glocal Description
1.5.3 Cross-Dimensional Indexing
1.5.4 Shape Embedding
1.5.5 Temporal Consistency
1.5.6 Experiments and Results
1.5.6.1 Quality
1.5.6.2 Scalability
1.5.7 Conclusions
1.6 Conclusions and Future Issues
Acknowledgment
Terminology Indexes
References
1.1 INTRODUCTION
The spread of digital multimedia content and services in the field of broadcasting and on the Internet has made multimedia data mining an important technology for transforming these sources into business intelligence for content owners, publishers, and distributors. A recent research domain known as multimedia duplicate mining (MDM) has emerged largely in response to this technological trend. The “multimedia duplicate mining” domain is based on detecting image, video, or audio copies from a test collection of multimedia resources. One very rich area of application is digital rights management, where the unauthorized or prohibited use of digital media on file-sharing networks can be detected to avoid copyright violations. The primary thesis of MDM in this application is “the media itself is the watermark,” that is, the media (image, video, or audio) contains enough unique information to be used to detect copies (Hampapur et al., 2002). The key advantage of MDM over other technologies, for example, the watermarking, is the fact that it can be introduced after copies are made and can be applied to content that is already in circulation.
Monitoring commercial films (CFs) is an important and valuable task for competitive marketing analysis, for advertising planning, and as a barometer of the advertising industry’s health in the field of market research (Li et al., 2005; Gauch and Shivadas, 2006; Herley, 2006; Berrani et al., 2008; Dohring and Lienhart, 2009; Putpuek et al., 2010; Wu and Satoh, 2011). In the field of broadcast media research, duplicate videos shared by multiple news programs imply that there are latent semantic relations between news stories. This information can be used to define the similarities between news stories; thus, it is useful for news story tracking, threading, and ranking (Duygulu et al., 2004; Zhai and Shah, 2005; Wu et al., 2008a; Wu et al., 2010). From another viewpoint, duplicate videos play a critical role in assessing the novelty and redundancy among news stories, and can help in identifying any fresh development among a huge volume of information in a limited amount of time (Wu et al., 2007a; Wu et al., 2008b). Additionally, MDM can be used to detect filler materials, for example, opening CG shots, anchor person shots, and weather charts in television, or background music in radio broadcasting (Satoh, 2002).
This chapter discusses the feasibility, techniques, and demonstrations of discovering hidden knowledge by applying MDM methods to the massive amount of multimedia content. We start by discussing the requirements and selection criteria for the duplicate mining methods in terms of the accuracy and scalability. These claims involve the sampling and description of videos, the indexing structure, and the retrieval process, which depend on the application purposes. We introduce three promising knowledge-discovery applications to show the benefits of duplicate mining. The first application (Wu and Satoh, 2011) is dedicated to fully unsupervised TV commercial mining for sociological analysis. It uses a dual-stage temporal recurrence hashing algorithm for ultra-fast detection of identical video sequences. The second application (Wu et al., 2010) focuses on news story retrieval and threading: it uses a one-to-one symmetric algorithm with a local interest point index structure to accurately detect identical news events. The third application (Poullot et al., 2008, Poullot et al., 2009) is for large-scale cross-domain video mining. It exploits any weak geometric consistencies between near-duplicate images and addresses the scalability issue of a near-duplicate search. Finally, a discussion on these techniques and applications is given.
1.2 SELECTION CRITERION OF DUPLICATE MINING METHODS
Choosing a duplicate mining method can be difficult because of the variety of methods currently available. The application provider must decide which method is best suited to their individual needs, and, of course, the type of duplicate that they want to use as the target. In this sense, the definition of a duplicate is generally subjective and, to a certain extent, does depend on the type of application being taken into consideration. The application ranges from exact duplicate mining, where no changes are allowed, to a more general definition that requires the resources to be of the same scene, but with possibly strong photometric or geometric transformations.
1.2.1 EXACT DUPLICATE MINING
One direction for MDM is to mine duplicate videos or audios derived from the original resource without any or with very few transformations. This type of duplicate is known as an exact duplicate or exact copy. Typical cases include TV CFs and file footages used in news. Most existing studies on exact duplicate mining introduce the concept of fingerprint or “hash” functions. This is a signal that has been extracted from each ...
